A Speech Feature Extraction Algorithm Based on the Nonlinear Hopf Cochlea Model

Abstract

The performance of speech recognition system is well or not is closely related to the characteristic parameters. For emulating human auditory system, a new method of speech feature extraction based on Hopf filter banks is presented. We modeled the extraction process of the MFCC, and used Hopf filter banks instead of the triangular filter banks. Then, we according the characteristics of Basilar Membranes in the cochlea to adjust the center frequency and bandwidth of the filter. The test speech goes through the Hopf filter banks, multi-dimensional eigenvectors will be obtained. After that, by Discrete Cosine Transformation, we will get the Hopf cepstral coefficients of the speech. Comparing with traditional feature MFCC, the speech recognition systems with Hopf characteristic parameters have better recognition rate and robustness characteristics in low Signal Noise Ratio (SNR) environment.